A Robot’s Successful Retry Does Not Prove It Diagnosed the Mistake

A simulated robot fails to lift or stack a cube, receives advice and succeeds on a later attempt. It is tempting to call that self-repair. But which part helped: another chance, a useful instruction, or information about the particular mistake?

A November 4, 2026 preprint by Siddharth Aphale and Ayushman Singh separates those possibilities. In its main experiment, concise task advice improved recovery. Advice chosen for the current failure, however, showed no detected advantage over a fixed task instruction or guidance carrying a different failure label.

That is not a finding that diagnosis never helps. It is a sharper distinction: finishing the task after a mistake does not, by itself, establish that the system understood what went wrong.

Another Attempt Helps. Useful Advice Helps More.

The researchers used a frozen Qwen2.5-Coder-7B model to write programs for a simulated Panda robot arm. “Frozen” means the model’s underlying learned parameters did not change. It received numerical information about the scene rather than camera images.

Every retry began from the original scene. This was not a robot trying to recover from a cube it had already knocked somewhere else. It was a fresh program attempting the reset task, with up to five retries after the first failure.

The main comparison used the same 252 initially failed programs on cube-lifting and stacking tasks. Without failure feedback, another round of code generation recovered 68.3% of those failures within five retries. A system that selected short instructions using a rule-based failure label recovered 85.3%. Its advantage also remained under matched total-token budgets, which limited the amount of text the model processed and generated.

Those percentages describe recovery from an already failed attempt, not the success rate of all jobs or the chance of succeeding on the very next try.

The advice could be straightforward: grasp the object before lifting it, use the documented robot commands, or lift after securing the object. Such instructions can improve a program without requiring the model to infer a new lesson from that particular failure.

Does the Advice Need to Match the Mistake?

To test that distinction, the researchers changed the information while keeping the retry opportunity.

One comparison supplied feedback from another episode of the same task, with a different failure label. Another reused one fixed, task-compatible instruction template across retries. Each recovered 86.1% of the initial failures, compared with 85.3% for feedback selected for the current failure.

The small differences do not prove that the approaches are equivalent. The confidence intervals still allow meaningful differences in either direction. But this experiment did not detect an advantage from matching the coarse failure label to the current mistake.

The mismatched advice was not random nonsense. It still concerned the same task and could contain useful instructions. That is precisely what makes the comparison informative: it asks whether the helpful ingredient was task guidance, rather than a correct diagnosis of this episode.

A separate experiment gave fixed guidance from the first attempt. Initial success rose from 19.3% to 36.9% across 700 new scenes. That comparison changed both the guidance and the instruction for generating the program, so it cannot isolate which change caused the improvement. Its overall advantage over failure-linked feedback after six executions was not statistically significant after the authors’ correction for multiple comparisons.

A Helpful Rule Can Become the Wrong Rule

The paper also shows why fixed advice is not a universal solution.

A nut-insertion task required grasping at the rim. The transferred feedback instead instructed the robot to grasp at the center. Across 40 initial failures, the feedback condition recovered 82.5%, while retries without that feedback recovered 97.5%.

The decline is consistent with the incompatible grasp instruction. A rule that helps in one task family can work against another. Better instructions must fit the job; adding more confident advice is not automatically an improvement.

These are bounded simulation results. Some scene configurations appeared in both development and evaluation, though the stored initial programs were different. The study used privileged scene information and coarse failure labels, not rich visual diagnosis, and it did not measure the full cost of developing its instruction templates.

Success and Diagnosis Need Different Evidence

A real robot recovering after a missed pick faces a further problem: the world may have changed. An object can move, a grip can slip, or the next action can make the situation worse. Resetting the simulator deliberately removes that part of recovery.

This study therefore does not settle how richer feedback or a larger model would perform on physical tasks. Its useful contribution is a set of comparisons: give the system another try, give it useful standing guidance, then ask what information about the actual failure adds.

Checking whether a robot really completed the task is one question. Establishing why its second attempt worked is another. A successful retry answers the first; a convincing self-repair claim needs evidence for the second.